On-device models
Dooze runs language models directly on your Mac. No server, no internet, no account required.
How it works
Dooze uses MLX to run models natively on Apple Silicon. Models are optimized for your hardware and run entirely in memory. Responses stay on your device.
Available models
Dooze ships with a small default model that works on most machines. You can download additional models in Settings > Models. Larger models produce better responses but need more RAM and respond slower.
RAM recommendations
| RAM | What works well |
|---|---|
| 8 GB | Small models, basic tasks |
| 16 GB | Medium models, good for most use |
| 32 GB+ | Large models, best quality |
Dooze shows estimated RAM usage before you download a model. If a model is too large for your machine, Dooze warns you.
Offline
On-device models work without an internet connection. Transcription, meeting recording, Chat, skills, all run locally. This is useful on planes, in restricted networks, or when you simply prefer not to send data anywhere.
Privacy
Nothing leaves your Mac. Your prompts, responses, transcriptions, and memories stay local. See Privacy.
Cloud models
When you want more capable responses, Lite and Heavy plans give you access to cloud models from OpenAI, Anthropic, xAI, and Google. You can switch between local and cloud models per conversation. See Models and tiers.